A Decision Model Breaks Like Any Other Language Model: A First Look at Jev
Check Point, Thursday, September 24th, 2026
Check Point found prompt injection could manipulate TypeSafe AI's Jev decision model in every configuration tested, for about 50 cents.
Check Point tested Jev, a new AI model from TypeSafe AI that returns typed decisions - like a score or yes/no - rather than text, meant to be consumed directly by software rather than read by people.
Using a due-diligence assistant scenario adapted from its Agent Breaker challenge, researchers found every configuration they tested could be manipulated, downgrading risk verdicts to low on documents flagging every warning sign, at a cost of about 50 cents per successful break and roughly four turns on average.
Structured, typed input and instructions to distrust the source document didn't meaningfully harden the model, and reasoning - the strongest defense Check Point measured - is not an available setting in Jev.